embeddings.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import os.path\n",
    "import math\n",
    "import re\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import facenet\n",
    "from sklearn.decomposition import PCA\n",
    "import matplotlib.pyplot as plt\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def get_model_filenames(model_dir):\n",
    "    files = os.listdir(model_dir)\n",
    "    meta_files = [s for s in files if s.endswith('.meta')]\n",
    "    if len(meta_files)==0:\n",
    "        raise ValueError('No meta file found in the model directory (%s)' % model_dir)\n",
    "    elif len(meta_files)>1:\n",
    "        raise ValueError('There should not be more than one meta file in the model directory (%s)' % model_dir)\n",
    "    meta_file = meta_files[0]\n",
    "    meta_files = [s for s in files if '.ckpt' in s]\n",
    "    max_step = -1\n",
    "    for f in files:\n",
    "        step_str = re.match(r'(^model-[\\w\\- ]+.ckpt-(\\d+))', f)\n",
    "        if step_str is not None and len(step_str.groups())>=2:\n",
    "            step = int(step_str.groups()[1])\n",
    "            if step > max_step:\n",
    "                max_step = step\n",
    "                ckpt_file = step_str.groups()[0]\n",
    "    return meta_file, ckpt_file\n",
    "\n",
    "def load_model(model, sess):\n",
    "    model_exp = os.path.expanduser(model)\n",
    "    print('Model directory: %s' % model_exp)\n",
    "    meta_file, ckpt_file = get_model_filenames(model_exp)\n",
    "\n",
    "    print('Metagraph file: %s' % meta_file)\n",
    "    print('Checkpoint file: %s' % ckpt_file)\n",
    "\n",
    "    saver = tf.train.import_meta_graph(os.path.join(model_exp, meta_file))\n",
    "    saver.restore(sess, os.path.join(model_exp, ckpt_file))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def forward(paths, batch_size):\n",
    "    nrof_images = len(paths)\n",
    "    nrof_batches = int(math.ceil(1.0*nrof_images / batch_size))\n",
    "    emb_array = np.zeros((nrof_images, embedding_size))\n",
    "\n",
    "    for i in range(nrof_batches):    \n",
    "        start_index = i*batch_size\n",
    "        end_index = min((i+1)*batch_size, nrof_images)\n",
    "        paths_batch = paths[start_index:end_index]\n",
    "        images = facenet.load_data(paths_batch, False, False, image_size)\n",
    "        feed_dict = { images_placeholder:images, phase_train_placeholder:False }\n",
    "        emb_array[start_index:end_index,:] = sess.run(embeddings, feed_dict=feed_dict)    \n",
    "        print(\"batch number:\", i)\n",
    "    print(\"Complete\")\n",
    "    \n",
    "    return emb_array\n",
    "\n",
    "def printDistance(emb_array, label_list, end=10):\n",
    "    threshold = 1.0\n",
    "    lng = len(emb_array[:end])\n",
    "    print('Distance matrix')\n",
    "    print('    ', end='')\n",
    "    for i in range(lng):\n",
    "        print('    %3d     ' % i, end='')\n",
    "    print('')\n",
    "    for i in range(lng):\n",
    "        print('%1d  ' % i, end='')\n",
    "        for j in range(lng):\n",
    "            dist = np.linalg.norm(emb_array[i,:] - emb_array[j,:])\n",
    "            print('  %1.4f' % dist, end='')\n",
    "            if (label_list[i] == label_list[j]) and (dist <= threshold):\n",
    "                print(\"(TP)\", end='')\n",
    "            elif (label_list[i] != label_list[j]) and (dist > threshold):\n",
    "                print(\"(TN)\", end='')\n",
    "            elif (label_list[i] == label_list[j]) and (dist > threshold):\n",
    "                print(\"\\033[93m(FN)\\033[0m\", end='')\n",
    "            elif (label_list[i] != label_list[j]) and (dist <= threshold):\n",
    "                print(\"\\033[93m(FP)\\033[0m\", end='')\n",
    "        print('')\n",
    "        \n",
    "def printData(labels, paths, number):\n",
    "    for i in range(number):\n",
    "        if i > 0 and labels[i] != labels[i-1]:\n",
    "            print('')\n",
    "        print('{0:3d} {1:3d}   {2:s}'.format(i, labels[i], paths[i].split(\"\\\\\")[1]))\n",
    "        # print('{0:3d}   {1:s}'.format(labels[i], paths[i].split(\"\\\\\")[1]))\n",
    "        \n",
    "def PCAtransform(emb_array, n_components = 10):\n",
    "    pca = PCA(n_components, whiten=True)\n",
    "    pca.fit(emb_array)\n",
    "    return pca.transform(emb_array), pca\n",
    "\n",
    "def splitTrainTest(data_set, percent_train):\n",
    "    paths, labels_list = facenet.get_image_paths_and_labels(data_set)\n",
    "    \n",
    "    lng = len(paths)\n",
    "    lng_train = int(lng * percent_train / 100);\n",
    "    lng_test = lng - lng_train\n",
    "    \n",
    "    train_paths = paths[0:lng_train]\n",
    "    test_paths = paths[lng_train:]\n",
    "    \n",
    "    train_labels = labels_list[0:lng_train]\n",
    "    test_labels = labels_list[lng_train:]\n",
    "    \n",
    "    return train_paths, train_labels, test_paths, test_labels\n",
    "\n",
    "def trainFisher(emb_array, y):\n",
    "    emb_del = np.empty_like(emb_array)\n",
    "    np.copyto(emb_del, emb_array)\n",
    "    \n",
    "    emb_array = np.vstack((emb_array, y))\n",
    "    \n",
    "    # y = emb_array[sample_number]\n",
    "    # emb_del = np.delete(emb_array, (sample_number), axis=0)\n",
    "    \n",
    "    E = np.linalg.inv(np.cov(emb_array.T))\n",
    "    D = (y - (1 / (len(emb_array) - 1) * np.sum(emb_del)))\n",
    "    \n",
    "    w = np.dot(E, D)\n",
    "    c = np.dot(w / np.linalg.norm(w), y)\n",
    "    \n",
    "    return c, w\n",
    "\n",
    "def trainFisherWithNum(emb_array, sample_number):    \n",
    "    y = emb_array[sample_number]\n",
    "    emb_del = np.delete(emb_array, (sample_number), axis=0)\n",
    "    \n",
    "    E = np.linalg.inv(np.cov(emb_array.T))\n",
    "    D = (y - (1 / (len(emb_array) - 1) * np.sum(emb_del)))\n",
    "    \n",
    "    w = np.dot(E, D)\n",
    "    c = np.dot(w / np.linalg.norm(w), y)\n",
    "    \n",
    "    return c, w    \n",
    "\n",
    "def Fisher(w, c, sample):\n",
    "    return np.dot(w / np.linalg.norm(w), sample) - c"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "sess = tf.Session()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "data_set = facenet.get_dataset(\"../../datasets/lfw/lfw_mtcnnpy_160/\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "train_paths, train_labels, test_paths, test_labels = splitTrainTest(data_set, 100)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model directory: 20170511-185253\n",
      "Metagraph file: model-20170511-185253.meta\n",
      "Checkpoint file: model-20170511-185253.ckpt-80000\n",
      "WARNING:tensorflow:The saved meta_graph is possibly from an older release:\n",
      "'model_variables' collection should be of type 'byte_list', but instead is of type 'node_list'.\n",
      "INFO:tensorflow:Restoring parameters from 20170511-185253\\model-20170511-185253.ckpt-80000\n"
     ]
    }
   ],
   "source": [
    "load_model(\"20170511-185253\", sess)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "images_placeholder = tf.get_default_graph().get_tensor_by_name(\"input:0\")\n",
    "embeddings = tf.get_default_graph().get_tensor_by_name(\"embeddings:0\")\n",
    "phase_train_placeholder = tf.get_default_graph().get_tensor_by_name(\"phase_train:0\")\n",
    "\n",
    "image_size = 160\n",
    "embedding_size = embeddings.get_shape()[1]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "batch number: 0\n",
      "batch number: 1\n",
      "batch number: 2\n",
      "batch number: 3\n",
      "batch number: 4\n",
      "batch number: 5\n",
      "batch number: 6\n",
      "batch number: 7\n",
      "batch number: 8\n",
      "batch number: 9\n",
      "batch number: 10\n",
      "batch number: 11\n",
      "batch number: 12\n",
      "batch number: 13\n",
      "batch number: 14\n",
      "batch number: 15\n",
      "batch number: 16\n",
      "batch number: 17\n",
      "batch number: 18\n",
      "batch number: 19\n",
      "batch number: 20\n",
      "batch number: 21\n",
      "batch number: 22\n",
      "batch number: 23\n",
      "batch number: 24\n",
      "batch number: 25\n",
      "batch number: 26\n",
      "batch number: 27\n",
      "batch number: 28\n",
      "batch number: 29\n",
      "batch number: 30\n",
      "batch number: 31\n",
      "batch number: 32\n",
      "batch number: 33\n",
      "batch number: 34\n",
      "batch number: 35\n",
      "batch number: 36\n",
      "batch number: 37\n",
      "batch number: 38\n",
      "batch number: 39\n",
      "batch number: 40\n",
      "batch number: 41\n",
      "batch number: 42\n",
      "batch number: 43\n",
      "batch number: 44\n",
      "batch number: 45\n",
      "batch number: 46\n",
      "batch number: 47\n",
      "batch number: 48\n",
      "batch number: 49\n",
      "batch number: 50\n",
      "batch number: 51\n",
      "batch number: 52\n",
      "batch number: 53\n",
      "batch number: 54\n",
      "batch number: 55\n",
      "batch number: 56\n",
      "batch number: 57\n",
      "batch number: 58\n",
      "batch number: 59\n",
      "batch number: 60\n",
      "batch number: 61\n",
      "batch number: 62\n",
      "batch number: 63\n",
      "batch number: 64\n",
      "batch number: 65\n",
      "batch number: 66\n",
      "batch number: 67\n",
      "batch number: 68\n",
      "batch number: 69\n",
      "batch number: 70\n",
      "batch number: 71\n",
      "batch number: 72\n",
      "batch number: 73\n",
      "batch number: 74\n",
      "batch number: 75\n",
      "batch number: 76\n",
      "batch number: 77\n",
      "batch number: 78\n",
      "batch number: 79\n",
      "batch number: 80\n",
      "batch number: 81\n",
      "batch number: 82\n",
      "batch number: 83\n",
      "batch number: 84\n",
      "batch number: 85\n",
      "batch number: 86\n",
      "batch number: 87\n",
      "batch number: 88\n",
      "batch number: 89\n",
      "batch number: 90\n",
      "batch number: 91\n",
      "batch number: 92\n",
      "batch number: 93\n",
      "batch number: 94\n",
      "batch number: 95\n",
      "batch number: 96\n",
      "batch number: 97\n",
      "batch number: 98\n",
      "batch number: 99\n",
      "batch number: 100\n",
      "batch number: 101\n",
      "batch number: 102\n",
      "batch number: 103\n",
      "batch number: 104\n",
      "batch number: 105\n",
      "batch number: 106\n",
      "batch number: 107\n",
      "batch number: 108\n",
      "batch number: 109\n",
      "batch number: 110\n",
      "batch number: 111\n",
      "batch number: 112\n",
      "batch number: 113\n",
      "batch number: 114\n",
      "batch number: 115\n",
      "batch number: 116\n",
      "batch number: 117\n",
      "batch number: 118\n",
      "batch number: 119\n",
      "batch number: 120\n",
      "batch number: 121\n",
      "batch number: 122\n",
      "batch number: 123\n",
      "batch number: 124\n",
      "batch number: 125\n",
      "batch number: 126\n",
      "batch number: 127\n",
      "batch number: 128\n",
      "batch number: 129\n",
      "batch number: 130\n",
      "batch number: 131\n",
      "batch number: 132\n",
      "Complete\n"
     ]
    }
   ],
   "source": [
    "emb_array = forward(train_paths, 100) # 100 means 100 mini-batches"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Centering data\n",
    "mean = np.mean(emb_array, axis=0)\n",
    "emb_array = emb_array - mean"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "emb_origin = np.empty_like(emb_array)\n",
    "np.copyto(emb_origin, emb_array)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "  0   0   AJ_Cook_0001.png\n",
      "\n",
      "  1   1   AJ_Lamas_0001.png\n",
      "\n",
      "  2   2   Aaron_Eckhart_0001.png\n",
      "\n",
      "  3   3   Aaron_Guiel_0001.png\n",
      "\n",
      "  4   4   Aaron_Patterson_0001.png\n",
      "\n",
      "  5   5   Aaron_Peirsol_0001.png\n",
      "  6   5   Aaron_Peirsol_0002.png\n",
      "  7   5   Aaron_Peirsol_0003.png\n",
      "  8   5   Aaron_Peirsol_0004.png\n",
      "\n",
      "  9   6   Aaron_Pena_0001.png\n",
      "\n",
      " 10   7   Aaron_Sorkin_0001.png\n",
      " 11   7   Aaron_Sorkin_0002.png\n",
      "\n",
      " 12   8   Aaron_Tippin_0001.png\n",
      "\n",
      " 13   9   Abba_Eban_0001.png\n",
      "\n",
      " 14  10   Abbas_Kiarostami_0001.png\n",
      "\n",
      " 15  11   Abdel_Aziz_Al-Hakim_0001.png\n",
      "\n",
      " 16  12   Abdel_Madi_Shabneh_0001.png\n",
      "\n",
      " 17  13   Abdel_Nasser_Assidi_0001.png\n",
      " 18  13   Abdel_Nasser_Assidi_0002.png\n",
      "\n",
      " 19  14   Abdoulaye_Wade_0001.png\n"
     ]
    }
   ],
   "source": [
    "printData(train_labels, train_paths, 20)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "emb_array = np.delete(emb_array, (6), axis=0)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAecAAAFpCAYAAACmt+D8AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0xfcd2c8e4e0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "pca = PCA().fit(emb_array)\n",
    "\n",
    "# 90% of variance\n",
    "plt.figure(figsize=(8, 6))\n",
    "plt.plot(np.cumsum(pca.explained_variance_ratio_), color='k', lw=2)\n",
    "plt.yticks(np.arange(0, 1.2, 0.1))\n",
    "plt.axvline(60, c='r')\n",
    "plt.axhline(1.0, c='r')\n",
    "plt.grid(b=True, which='major', color='grey', linestyle='-')\n",
    "plt.show()\n",
    "\n",
    "emb_pca, _ = PCAtransform(emb_array, 120)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "origin_pca, _ = PCAtransform(emb_origin, 120)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "c, w = trainFisherWithNum(emb_pca, 5)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "-0.111901597464\n",
      "-2.07136130855\n",
      "-3.24065270129\n",
      "-5.49479385434\n",
      "Accuracy: 1.0\n"
     ]
    }
   ],
   "source": [
    "lng = len(origin_pca)\n",
    "summ = 0\n",
    "thr = -5.5\n",
    "\n",
    "for i in range(lng):\n",
    "    l = Fisher(w, c, origin_pca[i])\n",
    "    if l > thr and train_labels[i] == 5:\n",
    "        summ += 1\n",
    "        print(l)\n",
    "    elif l < thr and train_labels[i] != 5:\n",
    "        summ += 1\n",
    "    else: print(i, l, train_labels[i])\n",
    "        \n",
    "print(\"Accuracy:\", summ / lng)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}